Sequoia Warns: The U.S. Won Closed-Source Models, but May Lose the Open Stack

According to Beating monitoring by a report, Sequoia partners Dean Meyer and Konstantine Buhler wrote that the US holds leading closed-source models, while Western companies are increasingly relying on China’s open-weight models. Qwen, Kimi, GLM, and DeepSeek are being used as product foundations, as training “teachers,” and as sources for synthetic data.

ATOM’s report shows that Qwen’s share in monthly newly fine-tuned and adapted models rose from 1% in January 2024 to 69% in February 2026. Thinking Machines’ Inkling also used synthetic data generated by open-weight models such as Kimi K2.5 in the early stages of post-training. However, this accounts for only a small portion of training compute.

The article argues the problem lies in “distillation” rules. Distillation is using a strong model’s outputs to train another model. While OpenAI and Anthropic restrict customers from using model outputs to train competing products, US companies can legally learn from China’s open models.

The two suggest that US cutting-edge labs sell qualified enterprises controlled, delayed, and auditable training rights. Otherwise, the US may continue to lead with closed-source models but hand the open-model foundation layer to China.

China has recently discussed limiting overseas access to certain advanced models, but no final policy has been formed yet. The article warns that even if current models can still be downloaded and used, Western companies may gradually fall behind due to not getting subsequent versions.